EDIC Framework Enables Per-Instruction Compliance for LLMs on External Data
October 19, 2026
The EDIC framework introduces per-instruction credit attribution to allow LLMs to follow legitimate directives from external data while rejecting adversarial injections. It uses a decomposed reward framework refined with GRPO to classify and act on individual instructions within a single document.
HOW THIS AFFECTS YOU
●
builderYou can implement more granular safety controls when feeding untrusted external documents into your RAG pipelines.
●
researcherThis provides a formal way to address the safety-utility tension in instruction-following models.